[linux-cifs-client] Re: extremely slow throughput writing to remote

[Steve French]

Steve French wrote:
> Just did a quick experiment:
>
> From my laptop to a home PC over standard 100Mbit ethernet (samba 
> 3.0.20 on 2.6.17-rc2 with ext3 on a common older IDE drive, WDC 1200JB).
>
> For a 256MB file copy
>     "dd if=/dev/zero of=/mnt/tstfile bs=256K count=1000"
> cifs vfs (2.6.17-rc2 on client) was 11.2 MB/sec (90 megabits per 
> seconds is near maximum wire speed)  vs. Local performance of large 
> file copy was over 30MB/sec.  This is in line with what I would expect 
> given the wire speed and much longer path lengths and lack of 
> serialization.  Note that with
>
> Note that the default block size of 512 that dd uses is atrocious and 
> causes cifs to be given very small writes - which is much slower 
> (2MB/sec).  Make sure your application is using a block size as least 
> as big as the cifs write size (14 pages - ie 56K)
>
> I will retry over gigE on the same two machines in a few minutes.
>
Tried 3 times over gigabit (same thing 256Megabytes: 256K write size, 
count 1000) and got more or less what I expected about 50% utilization - 
for the three test runs cifs to samba was averaging more than 55MB/sec.  
By comparison nfs (default server and client mount options) was less 
than 1/2 the speed of cifs (nfs averaged 22MB/sec).   IIRC nfs did much 
better (relative to cifs) on read than write.

Lots of room for improvement in cifs, but not bad.  I would like to ge 
it to about 90 MB/sec over gigabit.